There are many mysteries surrounding the finances of a major AI institute, and as Anthropic inches towards an IPO, we’re starting to get some answers. In particular, the company is finally providing details about the economics of computing and how much of Anthropic’s soaring revenue is being lost to computing costs.
In a new filing Monday, Anthropic released data showing that its compute margins are actually worsening. The filing was an update to the S-1 released last week, but the filing’s financial disclosures only covered June 30 of this year, but the new data extends additional quarterly earnings disclosures to September 30. Overall, Anthropic had a strong quarter. Revenue reached $17 billion, up from $11.5 billion in the previous quarter, suggesting sustained growth. However, operating loss also increased at the same time, mainly due to increased computing costs.
The situation gets even worse when you look at compute margins. In the new quarter, computing accounted for 58% of the company’s revenue, up from 48% in the previous quarter. The company may have been able to attract IPO investors by showing improved margins, but margins are deteriorating.
A graph in the filing shows this even more clearly, tracking the company’s revenue per dollar spent on computing. Since the beginning of 2025, that ratio has been steadily increasing, peaking at just over 2x this quarter, but falling to 1.72x since then.
While many signs suggest that this change in computing margins is more of a temporary issue than a long-term problem, it still tells us something important about the business of running major labs. In an effort to meet demand, Anthropic’s computing costs soared to $10 billion in the quarter. That includes, as the filing notes, “$5 billion in compute under short-term contracts, which is currently significantly higher than comparable computing costs procured under multi-year contracts.” Essentially, demand grew so quickly that Anthropic had to compete for compute, which meant paying higher prices.
Some of these deals are already being carried out in public. Anthropic pays SpaceX $1.25 billion a month for computing costs, on top of multibillion-dollar contracts with Nscale, Akamai, and Volta. In industry parlance, these are neoclouds rather than hyperscalers: smaller organizations offering more flexible computing deals at higher rates.
Anthropic’s margins suffered in the third quarter due to these deals, and there are no signs that the company will be able to get by without them any time soon. “As additional committed compute becomes available under our multi-year agreement with CSP and we continue to invest in additional compute, we expect our compute and infrastructure costs to increase in absolute terms and remain our largest operating expense for the foreseeable future,” the new filing states.
Anthropic is still working to reduce computing costs. In particular, the company has talked in considerable detail about its efforts to build its own computing base under a project it calls “Anthropic Direct.” However, as demand grows rapidly, these efforts cannot keep up, so Neocloud will continue to be the best option.
In a sense, these are champagne issues. As long as revenues are growing close to 50% per quarter, few investors will worry about high computing costs. And the moment demand slows, investors may take comfort in the brighter computing picture. But it’s a reminder of how dependent AI labs still are on multitrillion-dollar infrastructure developments, and how vulnerable they are to problems along the way. It’s easy to imagine data center construction slowing over the next year if bond markets become volatile or local opposition becomes more intense. There is not much Anthropic can do to address any of these risks. Both risks add to the messy computing picture.
In the meantime, the future of Anthropic and other institutes will be at the mercy of computing prices.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
